1 /* 2 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved. 3 * 4 * This software is available to you under a choice of one of two 5 * licenses. You may choose to be licensed under the terms of the GNU 6 * General Public License (GPL) Version 2, available from the file 7 * COPYING in the main directory of this source tree, or the 8 * OpenIB.org BSD license below: 9 * 10 * Redistribution and use in source and binary forms, with or 11 * without modification, are permitted provided that the following 12 * conditions are met: 13 * 14 * - Redistributions of source code must retain the above 15 * copyright notice, this list of conditions and the following 16 * disclaimer. 17 * 18 * - Redistributions in binary form must reproduce the above 19 * copyright notice, this list of conditions and the following 20 * disclaimer in the documentation and/or other materials 21 * provided with the distribution. 22 * 23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 30 * SOFTWARE. 31 */ 32 33 #include <linux/completion.h> 34 #include <linux/file.h> 35 #include <linux/mutex.h> 36 #include <linux/poll.h> 37 #include <linux/sched.h> 38 #include <linux/idr.h> 39 #include <linux/in.h> 40 #include <linux/in6.h> 41 #include <linux/miscdevice.h> 42 #include <linux/slab.h> 43 #include <linux/sysctl.h> 44 #include <linux/module.h> 45 46 #include <rdma/rdma_user_cm.h> 47 #include <rdma/ib_marshall.h> 48 #include <rdma/rdma_cm.h> 49 #include <rdma/rdma_cm_ib.h> 50 51 MODULE_AUTHOR("Sean Hefty"); 52 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access"); 53 MODULE_LICENSE("Dual BSD/GPL"); 54 55 static unsigned int max_backlog = 1024; 56 57 static struct ctl_table_header *ucma_ctl_table_hdr; 58 static ctl_table ucma_ctl_table[] = { 59 { 60 .procname = "max_backlog", 61 .data = &max_backlog, 62 .maxlen = sizeof max_backlog, 63 .mode = 0644, 64 .proc_handler = proc_dointvec, 65 }, 66 { } 67 }; 68 69 struct ucma_file { 70 struct mutex mut; 71 struct file *filp; 72 struct list_head ctx_list; 73 struct list_head event_list; 74 wait_queue_head_t poll_wait; 75 }; 76 77 struct ucma_context { 78 int id; 79 struct completion comp; 80 atomic_t ref; 81 int events_reported; 82 int backlog; 83 84 struct ucma_file *file; 85 struct rdma_cm_id *cm_id; 86 u64 uid; 87 88 struct list_head list; 89 struct list_head mc_list; 90 }; 91 92 struct ucma_multicast { 93 struct ucma_context *ctx; 94 int id; 95 int events_reported; 96 97 u64 uid; 98 struct list_head list; 99 struct sockaddr_storage addr; 100 }; 101 102 struct ucma_event { 103 struct ucma_context *ctx; 104 struct ucma_multicast *mc; 105 struct list_head list; 106 struct rdma_cm_id *cm_id; 107 struct rdma_ucm_event_resp resp; 108 }; 109 110 static DEFINE_MUTEX(mut); 111 static DEFINE_IDR(ctx_idr); 112 static DEFINE_IDR(multicast_idr); 113 114 static inline struct ucma_context *_ucma_find_context(int id, 115 struct ucma_file *file) 116 { 117 struct ucma_context *ctx; 118 119 ctx = idr_find(&ctx_idr, id); 120 if (!ctx) 121 ctx = ERR_PTR(-ENOENT); 122 else if (ctx->file != file) 123 ctx = ERR_PTR(-EINVAL); 124 return ctx; 125 } 126 127 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id) 128 { 129 struct ucma_context *ctx; 130 131 mutex_lock(&mut); 132 ctx = _ucma_find_context(id, file); 133 if (!IS_ERR(ctx)) 134 atomic_inc(&ctx->ref); 135 mutex_unlock(&mut); 136 return ctx; 137 } 138 139 static void ucma_put_ctx(struct ucma_context *ctx) 140 { 141 if (atomic_dec_and_test(&ctx->ref)) 142 complete(&ctx->comp); 143 } 144 145 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file) 146 { 147 struct ucma_context *ctx; 148 int ret; 149 150 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); 151 if (!ctx) 152 return NULL; 153 154 atomic_set(&ctx->ref, 1); 155 init_completion(&ctx->comp); 156 INIT_LIST_HEAD(&ctx->mc_list); 157 ctx->file = file; 158 159 do { 160 ret = idr_pre_get(&ctx_idr, GFP_KERNEL); 161 if (!ret) 162 goto error; 163 164 mutex_lock(&mut); 165 ret = idr_get_new(&ctx_idr, ctx, &ctx->id); 166 mutex_unlock(&mut); 167 } while (ret == -EAGAIN); 168 169 if (ret) 170 goto error; 171 172 list_add_tail(&ctx->list, &file->ctx_list); 173 return ctx; 174 175 error: 176 kfree(ctx); 177 return NULL; 178 } 179 180 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx) 181 { 182 struct ucma_multicast *mc; 183 int ret; 184 185 mc = kzalloc(sizeof(*mc), GFP_KERNEL); 186 if (!mc) 187 return NULL; 188 189 do { 190 ret = idr_pre_get(&multicast_idr, GFP_KERNEL); 191 if (!ret) 192 goto error; 193 194 mutex_lock(&mut); 195 ret = idr_get_new(&multicast_idr, mc, &mc->id); 196 mutex_unlock(&mut); 197 } while (ret == -EAGAIN); 198 199 if (ret) 200 goto error; 201 202 mc->ctx = ctx; 203 list_add_tail(&mc->list, &ctx->mc_list); 204 return mc; 205 206 error: 207 kfree(mc); 208 return NULL; 209 } 210 211 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst, 212 struct rdma_conn_param *src) 213 { 214 if (src->private_data_len) 215 memcpy(dst->private_data, src->private_data, 216 src->private_data_len); 217 dst->private_data_len = src->private_data_len; 218 dst->responder_resources =src->responder_resources; 219 dst->initiator_depth = src->initiator_depth; 220 dst->flow_control = src->flow_control; 221 dst->retry_count = src->retry_count; 222 dst->rnr_retry_count = src->rnr_retry_count; 223 dst->srq = src->srq; 224 dst->qp_num = src->qp_num; 225 } 226 227 static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst, 228 struct rdma_ud_param *src) 229 { 230 if (src->private_data_len) 231 memcpy(dst->private_data, src->private_data, 232 src->private_data_len); 233 dst->private_data_len = src->private_data_len; 234 ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr); 235 dst->qp_num = src->qp_num; 236 dst->qkey = src->qkey; 237 } 238 239 static void ucma_set_event_context(struct ucma_context *ctx, 240 struct rdma_cm_event *event, 241 struct ucma_event *uevent) 242 { 243 uevent->ctx = ctx; 244 switch (event->event) { 245 case RDMA_CM_EVENT_MULTICAST_JOIN: 246 case RDMA_CM_EVENT_MULTICAST_ERROR: 247 uevent->mc = (struct ucma_multicast *) 248 event->param.ud.private_data; 249 uevent->resp.uid = uevent->mc->uid; 250 uevent->resp.id = uevent->mc->id; 251 break; 252 default: 253 uevent->resp.uid = ctx->uid; 254 uevent->resp.id = ctx->id; 255 break; 256 } 257 } 258 259 static int ucma_event_handler(struct rdma_cm_id *cm_id, 260 struct rdma_cm_event *event) 261 { 262 struct ucma_event *uevent; 263 struct ucma_context *ctx = cm_id->context; 264 int ret = 0; 265 266 uevent = kzalloc(sizeof(*uevent), GFP_KERNEL); 267 if (!uevent) 268 return event->event == RDMA_CM_EVENT_CONNECT_REQUEST; 269 270 uevent->cm_id = cm_id; 271 ucma_set_event_context(ctx, event, uevent); 272 uevent->resp.event = event->event; 273 uevent->resp.status = event->status; 274 if (cm_id->qp_type == IB_QPT_UD) 275 ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud); 276 else 277 ucma_copy_conn_event(&uevent->resp.param.conn, 278 &event->param.conn); 279 280 mutex_lock(&ctx->file->mut); 281 if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) { 282 if (!ctx->backlog) { 283 ret = -ENOMEM; 284 kfree(uevent); 285 goto out; 286 } 287 ctx->backlog--; 288 } else if (!ctx->uid) { 289 /* 290 * We ignore events for new connections until userspace has set 291 * their context. This can only happen if an error occurs on a 292 * new connection before the user accepts it. This is okay, 293 * since the accept will just fail later. 294 */ 295 kfree(uevent); 296 goto out; 297 } 298 299 list_add_tail(&uevent->list, &ctx->file->event_list); 300 wake_up_interruptible(&ctx->file->poll_wait); 301 out: 302 mutex_unlock(&ctx->file->mut); 303 return ret; 304 } 305 306 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf, 307 int in_len, int out_len) 308 { 309 struct ucma_context *ctx; 310 struct rdma_ucm_get_event cmd; 311 struct ucma_event *uevent; 312 int ret = 0; 313 DEFINE_WAIT(wait); 314 315 if (out_len < sizeof uevent->resp) 316 return -ENOSPC; 317 318 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 319 return -EFAULT; 320 321 mutex_lock(&file->mut); 322 while (list_empty(&file->event_list)) { 323 mutex_unlock(&file->mut); 324 325 if (file->filp->f_flags & O_NONBLOCK) 326 return -EAGAIN; 327 328 if (wait_event_interruptible(file->poll_wait, 329 !list_empty(&file->event_list))) 330 return -ERESTARTSYS; 331 332 mutex_lock(&file->mut); 333 } 334 335 uevent = list_entry(file->event_list.next, struct ucma_event, list); 336 337 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) { 338 ctx = ucma_alloc_ctx(file); 339 if (!ctx) { 340 ret = -ENOMEM; 341 goto done; 342 } 343 uevent->ctx->backlog++; 344 ctx->cm_id = uevent->cm_id; 345 ctx->cm_id->context = ctx; 346 uevent->resp.id = ctx->id; 347 } 348 349 if (copy_to_user((void __user *)(unsigned long)cmd.response, 350 &uevent->resp, sizeof uevent->resp)) { 351 ret = -EFAULT; 352 goto done; 353 } 354 355 list_del(&uevent->list); 356 uevent->ctx->events_reported++; 357 if (uevent->mc) 358 uevent->mc->events_reported++; 359 kfree(uevent); 360 done: 361 mutex_unlock(&file->mut); 362 return ret; 363 } 364 365 static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type) 366 { 367 switch (cmd->ps) { 368 case RDMA_PS_TCP: 369 *qp_type = IB_QPT_RC; 370 return 0; 371 case RDMA_PS_UDP: 372 case RDMA_PS_IPOIB: 373 *qp_type = IB_QPT_UD; 374 return 0; 375 case RDMA_PS_IB: 376 *qp_type = cmd->qp_type; 377 return 0; 378 default: 379 return -EINVAL; 380 } 381 } 382 383 static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf, 384 int in_len, int out_len) 385 { 386 struct rdma_ucm_create_id cmd; 387 struct rdma_ucm_create_id_resp resp; 388 struct ucma_context *ctx; 389 enum ib_qp_type qp_type; 390 int ret; 391 392 if (out_len < sizeof(resp)) 393 return -ENOSPC; 394 395 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 396 return -EFAULT; 397 398 ret = ucma_get_qp_type(&cmd, &qp_type); 399 if (ret) 400 return ret; 401 402 mutex_lock(&file->mut); 403 ctx = ucma_alloc_ctx(file); 404 mutex_unlock(&file->mut); 405 if (!ctx) 406 return -ENOMEM; 407 408 ctx->uid = cmd.uid; 409 ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps, qp_type); 410 if (IS_ERR(ctx->cm_id)) { 411 ret = PTR_ERR(ctx->cm_id); 412 goto err1; 413 } 414 415 resp.id = ctx->id; 416 if (copy_to_user((void __user *)(unsigned long)cmd.response, 417 &resp, sizeof(resp))) { 418 ret = -EFAULT; 419 goto err2; 420 } 421 return 0; 422 423 err2: 424 rdma_destroy_id(ctx->cm_id); 425 err1: 426 mutex_lock(&mut); 427 idr_remove(&ctx_idr, ctx->id); 428 mutex_unlock(&mut); 429 kfree(ctx); 430 return ret; 431 } 432 433 static void ucma_cleanup_multicast(struct ucma_context *ctx) 434 { 435 struct ucma_multicast *mc, *tmp; 436 437 mutex_lock(&mut); 438 list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) { 439 list_del(&mc->list); 440 idr_remove(&multicast_idr, mc->id); 441 kfree(mc); 442 } 443 mutex_unlock(&mut); 444 } 445 446 static void ucma_cleanup_mc_events(struct ucma_multicast *mc) 447 { 448 struct ucma_event *uevent, *tmp; 449 450 list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) { 451 if (uevent->mc != mc) 452 continue; 453 454 list_del(&uevent->list); 455 kfree(uevent); 456 } 457 } 458 459 /* 460 * We cannot hold file->mut when calling rdma_destroy_id() or we can 461 * deadlock. We also acquire file->mut in ucma_event_handler(), and 462 * rdma_destroy_id() will wait until all callbacks have completed. 463 */ 464 static int ucma_free_ctx(struct ucma_context *ctx) 465 { 466 int events_reported; 467 struct ucma_event *uevent, *tmp; 468 LIST_HEAD(list); 469 470 /* No new events will be generated after destroying the id. */ 471 rdma_destroy_id(ctx->cm_id); 472 473 ucma_cleanup_multicast(ctx); 474 475 /* Cleanup events not yet reported to the user. */ 476 mutex_lock(&ctx->file->mut); 477 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) { 478 if (uevent->ctx == ctx) 479 list_move_tail(&uevent->list, &list); 480 } 481 list_del(&ctx->list); 482 mutex_unlock(&ctx->file->mut); 483 484 list_for_each_entry_safe(uevent, tmp, &list, list) { 485 list_del(&uevent->list); 486 if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) 487 rdma_destroy_id(uevent->cm_id); 488 kfree(uevent); 489 } 490 491 events_reported = ctx->events_reported; 492 kfree(ctx); 493 return events_reported; 494 } 495 496 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf, 497 int in_len, int out_len) 498 { 499 struct rdma_ucm_destroy_id cmd; 500 struct rdma_ucm_destroy_id_resp resp; 501 struct ucma_context *ctx; 502 int ret = 0; 503 504 if (out_len < sizeof(resp)) 505 return -ENOSPC; 506 507 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 508 return -EFAULT; 509 510 mutex_lock(&mut); 511 ctx = _ucma_find_context(cmd.id, file); 512 if (!IS_ERR(ctx)) 513 idr_remove(&ctx_idr, ctx->id); 514 mutex_unlock(&mut); 515 516 if (IS_ERR(ctx)) 517 return PTR_ERR(ctx); 518 519 ucma_put_ctx(ctx); 520 wait_for_completion(&ctx->comp); 521 resp.events_reported = ucma_free_ctx(ctx); 522 523 if (copy_to_user((void __user *)(unsigned long)cmd.response, 524 &resp, sizeof(resp))) 525 ret = -EFAULT; 526 527 return ret; 528 } 529 530 static ssize_t ucma_bind_addr(struct ucma_file *file, const char __user *inbuf, 531 int in_len, int out_len) 532 { 533 struct rdma_ucm_bind_addr cmd; 534 struct ucma_context *ctx; 535 int ret; 536 537 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 538 return -EFAULT; 539 540 ctx = ucma_get_ctx(file, cmd.id); 541 if (IS_ERR(ctx)) 542 return PTR_ERR(ctx); 543 544 ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr); 545 ucma_put_ctx(ctx); 546 return ret; 547 } 548 549 static ssize_t ucma_resolve_addr(struct ucma_file *file, 550 const char __user *inbuf, 551 int in_len, int out_len) 552 { 553 struct rdma_ucm_resolve_addr cmd; 554 struct ucma_context *ctx; 555 int ret; 556 557 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 558 return -EFAULT; 559 560 ctx = ucma_get_ctx(file, cmd.id); 561 if (IS_ERR(ctx)) 562 return PTR_ERR(ctx); 563 564 ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr, 565 (struct sockaddr *) &cmd.dst_addr, 566 cmd.timeout_ms); 567 ucma_put_ctx(ctx); 568 return ret; 569 } 570 571 static ssize_t ucma_resolve_route(struct ucma_file *file, 572 const char __user *inbuf, 573 int in_len, int out_len) 574 { 575 struct rdma_ucm_resolve_route cmd; 576 struct ucma_context *ctx; 577 int ret; 578 579 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 580 return -EFAULT; 581 582 ctx = ucma_get_ctx(file, cmd.id); 583 if (IS_ERR(ctx)) 584 return PTR_ERR(ctx); 585 586 ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms); 587 ucma_put_ctx(ctx); 588 return ret; 589 } 590 591 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp, 592 struct rdma_route *route) 593 { 594 struct rdma_dev_addr *dev_addr; 595 596 resp->num_paths = route->num_paths; 597 switch (route->num_paths) { 598 case 0: 599 dev_addr = &route->addr.dev_addr; 600 rdma_addr_get_dgid(dev_addr, 601 (union ib_gid *) &resp->ib_route[0].dgid); 602 rdma_addr_get_sgid(dev_addr, 603 (union ib_gid *) &resp->ib_route[0].sgid); 604 resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr)); 605 break; 606 case 2: 607 ib_copy_path_rec_to_user(&resp->ib_route[1], 608 &route->path_rec[1]); 609 /* fall through */ 610 case 1: 611 ib_copy_path_rec_to_user(&resp->ib_route[0], 612 &route->path_rec[0]); 613 break; 614 default: 615 break; 616 } 617 } 618 619 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp, 620 struct rdma_route *route) 621 { 622 struct rdma_dev_addr *dev_addr; 623 struct net_device *dev; 624 u16 vid = 0; 625 626 resp->num_paths = route->num_paths; 627 switch (route->num_paths) { 628 case 0: 629 dev_addr = &route->addr.dev_addr; 630 dev = dev_get_by_index(&init_net, dev_addr->bound_dev_if); 631 if (dev) { 632 vid = rdma_vlan_dev_vlan_id(dev); 633 dev_put(dev); 634 } 635 636 iboe_mac_vlan_to_ll((union ib_gid *) &resp->ib_route[0].dgid, 637 dev_addr->dst_dev_addr, vid); 638 iboe_addr_get_sgid(dev_addr, 639 (union ib_gid *) &resp->ib_route[0].sgid); 640 resp->ib_route[0].pkey = cpu_to_be16(0xffff); 641 break; 642 case 2: 643 ib_copy_path_rec_to_user(&resp->ib_route[1], 644 &route->path_rec[1]); 645 /* fall through */ 646 case 1: 647 ib_copy_path_rec_to_user(&resp->ib_route[0], 648 &route->path_rec[0]); 649 break; 650 default: 651 break; 652 } 653 } 654 655 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp, 656 struct rdma_route *route) 657 { 658 struct rdma_dev_addr *dev_addr; 659 660 dev_addr = &route->addr.dev_addr; 661 rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid); 662 rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid); 663 } 664 665 static ssize_t ucma_query_route(struct ucma_file *file, 666 const char __user *inbuf, 667 int in_len, int out_len) 668 { 669 struct rdma_ucm_query_route cmd; 670 struct rdma_ucm_query_route_resp resp; 671 struct ucma_context *ctx; 672 struct sockaddr *addr; 673 int ret = 0; 674 675 if (out_len < sizeof(resp)) 676 return -ENOSPC; 677 678 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 679 return -EFAULT; 680 681 ctx = ucma_get_ctx(file, cmd.id); 682 if (IS_ERR(ctx)) 683 return PTR_ERR(ctx); 684 685 memset(&resp, 0, sizeof resp); 686 addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr; 687 memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ? 688 sizeof(struct sockaddr_in) : 689 sizeof(struct sockaddr_in6)); 690 addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr; 691 memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ? 692 sizeof(struct sockaddr_in) : 693 sizeof(struct sockaddr_in6)); 694 if (!ctx->cm_id->device) 695 goto out; 696 697 resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid; 698 resp.port_num = ctx->cm_id->port_num; 699 switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) { 700 case RDMA_TRANSPORT_IB: 701 switch (rdma_port_get_link_layer(ctx->cm_id->device, 702 ctx->cm_id->port_num)) { 703 case IB_LINK_LAYER_INFINIBAND: 704 ucma_copy_ib_route(&resp, &ctx->cm_id->route); 705 break; 706 case IB_LINK_LAYER_ETHERNET: 707 ucma_copy_iboe_route(&resp, &ctx->cm_id->route); 708 break; 709 default: 710 break; 711 } 712 break; 713 case RDMA_TRANSPORT_IWARP: 714 ucma_copy_iw_route(&resp, &ctx->cm_id->route); 715 break; 716 default: 717 break; 718 } 719 720 out: 721 if (copy_to_user((void __user *)(unsigned long)cmd.response, 722 &resp, sizeof(resp))) 723 ret = -EFAULT; 724 725 ucma_put_ctx(ctx); 726 return ret; 727 } 728 729 static void ucma_copy_conn_param(struct rdma_conn_param *dst, 730 struct rdma_ucm_conn_param *src) 731 { 732 dst->private_data = src->private_data; 733 dst->private_data_len = src->private_data_len; 734 dst->responder_resources =src->responder_resources; 735 dst->initiator_depth = src->initiator_depth; 736 dst->flow_control = src->flow_control; 737 dst->retry_count = src->retry_count; 738 dst->rnr_retry_count = src->rnr_retry_count; 739 dst->srq = src->srq; 740 dst->qp_num = src->qp_num; 741 } 742 743 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf, 744 int in_len, int out_len) 745 { 746 struct rdma_ucm_connect cmd; 747 struct rdma_conn_param conn_param; 748 struct ucma_context *ctx; 749 int ret; 750 751 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 752 return -EFAULT; 753 754 if (!cmd.conn_param.valid) 755 return -EINVAL; 756 757 ctx = ucma_get_ctx(file, cmd.id); 758 if (IS_ERR(ctx)) 759 return PTR_ERR(ctx); 760 761 ucma_copy_conn_param(&conn_param, &cmd.conn_param); 762 ret = rdma_connect(ctx->cm_id, &conn_param); 763 ucma_put_ctx(ctx); 764 return ret; 765 } 766 767 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf, 768 int in_len, int out_len) 769 { 770 struct rdma_ucm_listen cmd; 771 struct ucma_context *ctx; 772 int ret; 773 774 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 775 return -EFAULT; 776 777 ctx = ucma_get_ctx(file, cmd.id); 778 if (IS_ERR(ctx)) 779 return PTR_ERR(ctx); 780 781 ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ? 782 cmd.backlog : max_backlog; 783 ret = rdma_listen(ctx->cm_id, ctx->backlog); 784 ucma_put_ctx(ctx); 785 return ret; 786 } 787 788 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf, 789 int in_len, int out_len) 790 { 791 struct rdma_ucm_accept cmd; 792 struct rdma_conn_param conn_param; 793 struct ucma_context *ctx; 794 int ret; 795 796 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 797 return -EFAULT; 798 799 ctx = ucma_get_ctx(file, cmd.id); 800 if (IS_ERR(ctx)) 801 return PTR_ERR(ctx); 802 803 if (cmd.conn_param.valid) { 804 ucma_copy_conn_param(&conn_param, &cmd.conn_param); 805 mutex_lock(&file->mut); 806 ret = rdma_accept(ctx->cm_id, &conn_param); 807 if (!ret) 808 ctx->uid = cmd.uid; 809 mutex_unlock(&file->mut); 810 } else 811 ret = rdma_accept(ctx->cm_id, NULL); 812 813 ucma_put_ctx(ctx); 814 return ret; 815 } 816 817 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf, 818 int in_len, int out_len) 819 { 820 struct rdma_ucm_reject cmd; 821 struct ucma_context *ctx; 822 int ret; 823 824 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 825 return -EFAULT; 826 827 ctx = ucma_get_ctx(file, cmd.id); 828 if (IS_ERR(ctx)) 829 return PTR_ERR(ctx); 830 831 ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len); 832 ucma_put_ctx(ctx); 833 return ret; 834 } 835 836 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf, 837 int in_len, int out_len) 838 { 839 struct rdma_ucm_disconnect cmd; 840 struct ucma_context *ctx; 841 int ret; 842 843 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 844 return -EFAULT; 845 846 ctx = ucma_get_ctx(file, cmd.id); 847 if (IS_ERR(ctx)) 848 return PTR_ERR(ctx); 849 850 ret = rdma_disconnect(ctx->cm_id); 851 ucma_put_ctx(ctx); 852 return ret; 853 } 854 855 static ssize_t ucma_init_qp_attr(struct ucma_file *file, 856 const char __user *inbuf, 857 int in_len, int out_len) 858 { 859 struct rdma_ucm_init_qp_attr cmd; 860 struct ib_uverbs_qp_attr resp; 861 struct ucma_context *ctx; 862 struct ib_qp_attr qp_attr; 863 int ret; 864 865 if (out_len < sizeof(resp)) 866 return -ENOSPC; 867 868 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 869 return -EFAULT; 870 871 ctx = ucma_get_ctx(file, cmd.id); 872 if (IS_ERR(ctx)) 873 return PTR_ERR(ctx); 874 875 resp.qp_attr_mask = 0; 876 memset(&qp_attr, 0, sizeof qp_attr); 877 qp_attr.qp_state = cmd.qp_state; 878 ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask); 879 if (ret) 880 goto out; 881 882 ib_copy_qp_attr_to_user(&resp, &qp_attr); 883 if (copy_to_user((void __user *)(unsigned long)cmd.response, 884 &resp, sizeof(resp))) 885 ret = -EFAULT; 886 887 out: 888 ucma_put_ctx(ctx); 889 return ret; 890 } 891 892 static int ucma_set_option_id(struct ucma_context *ctx, int optname, 893 void *optval, size_t optlen) 894 { 895 int ret = 0; 896 897 switch (optname) { 898 case RDMA_OPTION_ID_TOS: 899 if (optlen != sizeof(u8)) { 900 ret = -EINVAL; 901 break; 902 } 903 rdma_set_service_type(ctx->cm_id, *((u8 *) optval)); 904 break; 905 case RDMA_OPTION_ID_REUSEADDR: 906 if (optlen != sizeof(int)) { 907 ret = -EINVAL; 908 break; 909 } 910 ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0); 911 break; 912 default: 913 ret = -ENOSYS; 914 } 915 916 return ret; 917 } 918 919 static int ucma_set_ib_path(struct ucma_context *ctx, 920 struct ib_path_rec_data *path_data, size_t optlen) 921 { 922 struct ib_sa_path_rec sa_path; 923 struct rdma_cm_event event; 924 int ret; 925 926 if (optlen % sizeof(*path_data)) 927 return -EINVAL; 928 929 for (; optlen; optlen -= sizeof(*path_data), path_data++) { 930 if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY | 931 IB_PATH_BIDIRECTIONAL)) 932 break; 933 } 934 935 if (!optlen) 936 return -EINVAL; 937 938 ib_sa_unpack_path(path_data->path_rec, &sa_path); 939 ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1); 940 if (ret) 941 return ret; 942 943 memset(&event, 0, sizeof event); 944 event.event = RDMA_CM_EVENT_ROUTE_RESOLVED; 945 return ucma_event_handler(ctx->cm_id, &event); 946 } 947 948 static int ucma_set_option_ib(struct ucma_context *ctx, int optname, 949 void *optval, size_t optlen) 950 { 951 int ret; 952 953 switch (optname) { 954 case RDMA_OPTION_IB_PATH: 955 ret = ucma_set_ib_path(ctx, optval, optlen); 956 break; 957 default: 958 ret = -ENOSYS; 959 } 960 961 return ret; 962 } 963 964 static int ucma_set_option_level(struct ucma_context *ctx, int level, 965 int optname, void *optval, size_t optlen) 966 { 967 int ret; 968 969 switch (level) { 970 case RDMA_OPTION_ID: 971 ret = ucma_set_option_id(ctx, optname, optval, optlen); 972 break; 973 case RDMA_OPTION_IB: 974 ret = ucma_set_option_ib(ctx, optname, optval, optlen); 975 break; 976 default: 977 ret = -ENOSYS; 978 } 979 980 return ret; 981 } 982 983 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf, 984 int in_len, int out_len) 985 { 986 struct rdma_ucm_set_option cmd; 987 struct ucma_context *ctx; 988 void *optval; 989 int ret; 990 991 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 992 return -EFAULT; 993 994 ctx = ucma_get_ctx(file, cmd.id); 995 if (IS_ERR(ctx)) 996 return PTR_ERR(ctx); 997 998 optval = kmalloc(cmd.optlen, GFP_KERNEL); 999 if (!optval) { 1000 ret = -ENOMEM; 1001 goto out1; 1002 } 1003 1004 if (copy_from_user(optval, (void __user *) (unsigned long) cmd.optval, 1005 cmd.optlen)) { 1006 ret = -EFAULT; 1007 goto out2; 1008 } 1009 1010 ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval, 1011 cmd.optlen); 1012 out2: 1013 kfree(optval); 1014 out1: 1015 ucma_put_ctx(ctx); 1016 return ret; 1017 } 1018 1019 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf, 1020 int in_len, int out_len) 1021 { 1022 struct rdma_ucm_notify cmd; 1023 struct ucma_context *ctx; 1024 int ret; 1025 1026 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1027 return -EFAULT; 1028 1029 ctx = ucma_get_ctx(file, cmd.id); 1030 if (IS_ERR(ctx)) 1031 return PTR_ERR(ctx); 1032 1033 ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event); 1034 ucma_put_ctx(ctx); 1035 return ret; 1036 } 1037 1038 static ssize_t ucma_join_multicast(struct ucma_file *file, 1039 const char __user *inbuf, 1040 int in_len, int out_len) 1041 { 1042 struct rdma_ucm_join_mcast cmd; 1043 struct rdma_ucm_create_id_resp resp; 1044 struct ucma_context *ctx; 1045 struct ucma_multicast *mc; 1046 int ret; 1047 1048 if (out_len < sizeof(resp)) 1049 return -ENOSPC; 1050 1051 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1052 return -EFAULT; 1053 1054 ctx = ucma_get_ctx(file, cmd.id); 1055 if (IS_ERR(ctx)) 1056 return PTR_ERR(ctx); 1057 1058 mutex_lock(&file->mut); 1059 mc = ucma_alloc_multicast(ctx); 1060 if (!mc) { 1061 ret = -ENOMEM; 1062 goto err1; 1063 } 1064 1065 mc->uid = cmd.uid; 1066 memcpy(&mc->addr, &cmd.addr, sizeof cmd.addr); 1067 ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc); 1068 if (ret) 1069 goto err2; 1070 1071 resp.id = mc->id; 1072 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1073 &resp, sizeof(resp))) { 1074 ret = -EFAULT; 1075 goto err3; 1076 } 1077 1078 mutex_unlock(&file->mut); 1079 ucma_put_ctx(ctx); 1080 return 0; 1081 1082 err3: 1083 rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr); 1084 ucma_cleanup_mc_events(mc); 1085 err2: 1086 mutex_lock(&mut); 1087 idr_remove(&multicast_idr, mc->id); 1088 mutex_unlock(&mut); 1089 list_del(&mc->list); 1090 kfree(mc); 1091 err1: 1092 mutex_unlock(&file->mut); 1093 ucma_put_ctx(ctx); 1094 return ret; 1095 } 1096 1097 static ssize_t ucma_leave_multicast(struct ucma_file *file, 1098 const char __user *inbuf, 1099 int in_len, int out_len) 1100 { 1101 struct rdma_ucm_destroy_id cmd; 1102 struct rdma_ucm_destroy_id_resp resp; 1103 struct ucma_multicast *mc; 1104 int ret = 0; 1105 1106 if (out_len < sizeof(resp)) 1107 return -ENOSPC; 1108 1109 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1110 return -EFAULT; 1111 1112 mutex_lock(&mut); 1113 mc = idr_find(&multicast_idr, cmd.id); 1114 if (!mc) 1115 mc = ERR_PTR(-ENOENT); 1116 else if (mc->ctx->file != file) 1117 mc = ERR_PTR(-EINVAL); 1118 else { 1119 idr_remove(&multicast_idr, mc->id); 1120 atomic_inc(&mc->ctx->ref); 1121 } 1122 mutex_unlock(&mut); 1123 1124 if (IS_ERR(mc)) { 1125 ret = PTR_ERR(mc); 1126 goto out; 1127 } 1128 1129 rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr); 1130 mutex_lock(&mc->ctx->file->mut); 1131 ucma_cleanup_mc_events(mc); 1132 list_del(&mc->list); 1133 mutex_unlock(&mc->ctx->file->mut); 1134 1135 ucma_put_ctx(mc->ctx); 1136 resp.events_reported = mc->events_reported; 1137 kfree(mc); 1138 1139 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1140 &resp, sizeof(resp))) 1141 ret = -EFAULT; 1142 out: 1143 return ret; 1144 } 1145 1146 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2) 1147 { 1148 /* Acquire mutex's based on pointer comparison to prevent deadlock. */ 1149 if (file1 < file2) { 1150 mutex_lock(&file1->mut); 1151 mutex_lock(&file2->mut); 1152 } else { 1153 mutex_lock(&file2->mut); 1154 mutex_lock(&file1->mut); 1155 } 1156 } 1157 1158 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2) 1159 { 1160 if (file1 < file2) { 1161 mutex_unlock(&file2->mut); 1162 mutex_unlock(&file1->mut); 1163 } else { 1164 mutex_unlock(&file1->mut); 1165 mutex_unlock(&file2->mut); 1166 } 1167 } 1168 1169 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file) 1170 { 1171 struct ucma_event *uevent, *tmp; 1172 1173 list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) 1174 if (uevent->ctx == ctx) 1175 list_move_tail(&uevent->list, &file->event_list); 1176 } 1177 1178 static ssize_t ucma_migrate_id(struct ucma_file *new_file, 1179 const char __user *inbuf, 1180 int in_len, int out_len) 1181 { 1182 struct rdma_ucm_migrate_id cmd; 1183 struct rdma_ucm_migrate_resp resp; 1184 struct ucma_context *ctx; 1185 struct file *filp; 1186 struct ucma_file *cur_file; 1187 int ret = 0; 1188 1189 if (copy_from_user(&cmd, inbuf, sizeof(cmd))) 1190 return -EFAULT; 1191 1192 /* Get current fd to protect against it being closed */ 1193 filp = fget(cmd.fd); 1194 if (!filp) 1195 return -ENOENT; 1196 1197 /* Validate current fd and prevent destruction of id. */ 1198 ctx = ucma_get_ctx(filp->private_data, cmd.id); 1199 if (IS_ERR(ctx)) { 1200 ret = PTR_ERR(ctx); 1201 goto file_put; 1202 } 1203 1204 cur_file = ctx->file; 1205 if (cur_file == new_file) { 1206 resp.events_reported = ctx->events_reported; 1207 goto response; 1208 } 1209 1210 /* 1211 * Migrate events between fd's, maintaining order, and avoiding new 1212 * events being added before existing events. 1213 */ 1214 ucma_lock_files(cur_file, new_file); 1215 mutex_lock(&mut); 1216 1217 list_move_tail(&ctx->list, &new_file->ctx_list); 1218 ucma_move_events(ctx, new_file); 1219 ctx->file = new_file; 1220 resp.events_reported = ctx->events_reported; 1221 1222 mutex_unlock(&mut); 1223 ucma_unlock_files(cur_file, new_file); 1224 1225 response: 1226 if (copy_to_user((void __user *)(unsigned long)cmd.response, 1227 &resp, sizeof(resp))) 1228 ret = -EFAULT; 1229 1230 ucma_put_ctx(ctx); 1231 file_put: 1232 fput(filp); 1233 return ret; 1234 } 1235 1236 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file, 1237 const char __user *inbuf, 1238 int in_len, int out_len) = { 1239 [RDMA_USER_CM_CMD_CREATE_ID] = ucma_create_id, 1240 [RDMA_USER_CM_CMD_DESTROY_ID] = ucma_destroy_id, 1241 [RDMA_USER_CM_CMD_BIND_ADDR] = ucma_bind_addr, 1242 [RDMA_USER_CM_CMD_RESOLVE_ADDR] = ucma_resolve_addr, 1243 [RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route, 1244 [RDMA_USER_CM_CMD_QUERY_ROUTE] = ucma_query_route, 1245 [RDMA_USER_CM_CMD_CONNECT] = ucma_connect, 1246 [RDMA_USER_CM_CMD_LISTEN] = ucma_listen, 1247 [RDMA_USER_CM_CMD_ACCEPT] = ucma_accept, 1248 [RDMA_USER_CM_CMD_REJECT] = ucma_reject, 1249 [RDMA_USER_CM_CMD_DISCONNECT] = ucma_disconnect, 1250 [RDMA_USER_CM_CMD_INIT_QP_ATTR] = ucma_init_qp_attr, 1251 [RDMA_USER_CM_CMD_GET_EVENT] = ucma_get_event, 1252 [RDMA_USER_CM_CMD_GET_OPTION] = NULL, 1253 [RDMA_USER_CM_CMD_SET_OPTION] = ucma_set_option, 1254 [RDMA_USER_CM_CMD_NOTIFY] = ucma_notify, 1255 [RDMA_USER_CM_CMD_JOIN_MCAST] = ucma_join_multicast, 1256 [RDMA_USER_CM_CMD_LEAVE_MCAST] = ucma_leave_multicast, 1257 [RDMA_USER_CM_CMD_MIGRATE_ID] = ucma_migrate_id 1258 }; 1259 1260 static ssize_t ucma_write(struct file *filp, const char __user *buf, 1261 size_t len, loff_t *pos) 1262 { 1263 struct ucma_file *file = filp->private_data; 1264 struct rdma_ucm_cmd_hdr hdr; 1265 ssize_t ret; 1266 1267 if (len < sizeof(hdr)) 1268 return -EINVAL; 1269 1270 if (copy_from_user(&hdr, buf, sizeof(hdr))) 1271 return -EFAULT; 1272 1273 if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table)) 1274 return -EINVAL; 1275 1276 if (hdr.in + sizeof(hdr) > len) 1277 return -EINVAL; 1278 1279 if (!ucma_cmd_table[hdr.cmd]) 1280 return -ENOSYS; 1281 1282 ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out); 1283 if (!ret) 1284 ret = len; 1285 1286 return ret; 1287 } 1288 1289 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait) 1290 { 1291 struct ucma_file *file = filp->private_data; 1292 unsigned int mask = 0; 1293 1294 poll_wait(filp, &file->poll_wait, wait); 1295 1296 if (!list_empty(&file->event_list)) 1297 mask = POLLIN | POLLRDNORM; 1298 1299 return mask; 1300 } 1301 1302 /* 1303 * ucma_open() does not need the BKL: 1304 * 1305 * - no global state is referred to; 1306 * - there is no ioctl method to race against; 1307 * - no further module initialization is required for open to work 1308 * after the device is registered. 1309 */ 1310 static int ucma_open(struct inode *inode, struct file *filp) 1311 { 1312 struct ucma_file *file; 1313 1314 file = kmalloc(sizeof *file, GFP_KERNEL); 1315 if (!file) 1316 return -ENOMEM; 1317 1318 INIT_LIST_HEAD(&file->event_list); 1319 INIT_LIST_HEAD(&file->ctx_list); 1320 init_waitqueue_head(&file->poll_wait); 1321 mutex_init(&file->mut); 1322 1323 filp->private_data = file; 1324 file->filp = filp; 1325 1326 return nonseekable_open(inode, filp); 1327 } 1328 1329 static int ucma_close(struct inode *inode, struct file *filp) 1330 { 1331 struct ucma_file *file = filp->private_data; 1332 struct ucma_context *ctx, *tmp; 1333 1334 mutex_lock(&file->mut); 1335 list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) { 1336 mutex_unlock(&file->mut); 1337 1338 mutex_lock(&mut); 1339 idr_remove(&ctx_idr, ctx->id); 1340 mutex_unlock(&mut); 1341 1342 ucma_free_ctx(ctx); 1343 mutex_lock(&file->mut); 1344 } 1345 mutex_unlock(&file->mut); 1346 kfree(file); 1347 return 0; 1348 } 1349 1350 static const struct file_operations ucma_fops = { 1351 .owner = THIS_MODULE, 1352 .open = ucma_open, 1353 .release = ucma_close, 1354 .write = ucma_write, 1355 .poll = ucma_poll, 1356 .llseek = no_llseek, 1357 }; 1358 1359 static struct miscdevice ucma_misc = { 1360 .minor = MISC_DYNAMIC_MINOR, 1361 .name = "rdma_cm", 1362 .nodename = "infiniband/rdma_cm", 1363 .mode = 0666, 1364 .fops = &ucma_fops, 1365 }; 1366 1367 static ssize_t show_abi_version(struct device *dev, 1368 struct device_attribute *attr, 1369 char *buf) 1370 { 1371 return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION); 1372 } 1373 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL); 1374 1375 static int __init ucma_init(void) 1376 { 1377 int ret; 1378 1379 ret = misc_register(&ucma_misc); 1380 if (ret) 1381 return ret; 1382 1383 ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version); 1384 if (ret) { 1385 printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n"); 1386 goto err1; 1387 } 1388 1389 ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table); 1390 if (!ucma_ctl_table_hdr) { 1391 printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n"); 1392 ret = -ENOMEM; 1393 goto err2; 1394 } 1395 return 0; 1396 err2: 1397 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version); 1398 err1: 1399 misc_deregister(&ucma_misc); 1400 return ret; 1401 } 1402 1403 static void __exit ucma_cleanup(void) 1404 { 1405 unregister_net_sysctl_table(ucma_ctl_table_hdr); 1406 device_remove_file(ucma_misc.this_device, &dev_attr_abi_version); 1407 misc_deregister(&ucma_misc); 1408 idr_destroy(&ctx_idr); 1409 } 1410 1411 module_init(ucma_init); 1412 module_exit(ucma_cleanup); 1413